On 2022/7/14 12:51, Barry Song wrote: > On Thu, Jul 14, 2022 at 3:29 PM Xin Hao <xhao@xxxxxxxxxxxxxxxxx> wrote: >> >> Hi barry. >> >> I do some test on Kunpeng arm64 machine use Unixbench. >> >> The test result as below. >> >> One core, we can see the performance improvement above +30%. > > I am really pleased to see the 30%+ improvement on unixbench on single core. > >> ./Run -c 1 -i 1 shell1 >> w/o >> System Benchmarks Partial Index BASELINE RESULT INDEX >> Shell Scripts (1 concurrent) 42.4 5481.0 1292.7 >> ======== >> System Benchmarks Index Score (Partial Only) 1292.7 >> >> w/ >> System Benchmarks Partial Index BASELINE RESULT INDEX >> Shell Scripts (1 concurrent) 42.4 6974.6 1645.0 >> ======== >> System Benchmarks Index Score (Partial Only) 1645.0 >> >> >> But with whole cores, there have little performance degradation above -5% > > That is sad as we might get more concurrency between mprotect(), madvise(), > mremap(), zap_pte_range() and the deferred tlbi. > >> >> ./Run -c 96 -i 1 shell1 >> w/o >> Shell Scripts (1 concurrent) 80765.5 lpm (60.0 s, 1 >> samples) >> System Benchmarks Partial Index BASELINE RESULT INDEX >> Shell Scripts (1 concurrent) 42.4 80765.5 19048.5 >> ======== >> System Benchmarks Index Score (Partial Only) 19048.5 >> >> w >> Shell Scripts (1 concurrent) 76333.6 lpm (60.0 s, 1 >> samples) >> System Benchmarks Partial Index BASELINE RESULT INDEX >> Shell Scripts (1 concurrent) 42.4 76333.6 18003.2 >> ======== >> System Benchmarks Index Score (Partial Only) 18003.2 >> >> ---------------------------------------------------------------------------------------------- >> >> >> After discuss with you, and do some changes in the patch. >> >> ndex a52381a680db..1ecba81f1277 100644 >> --- a/mm/rmap.c >> +++ b/mm/rmap.c >> @@ -727,7 +727,11 @@ void flush_tlb_batched_pending(struct mm_struct *mm) >> int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT; >> >> if (pending != flushed) { >> +#ifdef CONFIG_ARCH_HAS_MM_CPUMASK >> flush_tlb_mm(mm); >> +#else >> + dsb(ish); >> +#endif >> > > i was guessing the problem might be flush_tlb_batched_pending() > so i asked you to change this to verify my guess. > flush_tlb_batched_pending() looks like the critical path for this issue then the code above can mitigate this. I cannot reproduce this on a 2P 128C Kunpeng920 server. The kernel is based on the v5.19-rc6 and unixbench of version 5.1.3. The result of `./Run -c 128 -i 1 shell1` is: iter-1 iter-2 iter-3 w/o 17708.1 17637.1 17630.1 w 17766.0 17752.3 17861.7 And flush_tlb_batched_pending()isn't the hot spot with the patch: 7.00% sh [kernel.kallsyms] [k] ptep_clear_flush 4.17% sh [kernel.kallsyms] [k] ptep_set_access_flags 2.43% multi.sh [kernel.kallsyms] [k] ptep_clear_flush 1.98% sh [kernel.kallsyms] [k] _raw_spin_unlock_irqrestore 1.69% sh [kernel.kallsyms] [k] next_uptodate_page 1.66% sort [kernel.kallsyms] [k] ptep_clear_flush 1.56% multi.sh [kernel.kallsyms] [k] ptep_set_access_flags 1.27% sh [kernel.kallsyms] [k] page_counter_cancel 1.11% sh [kernel.kallsyms] [k] page_remove_rmap 1.06% sh [kernel.kallsyms] [k] perf_event_alloc Hi Xin Hao, I'm not sure the test setup as well as the config is same with yours. (96C vs 128C should not be the reason I think). Did you check that the 5% is a fluctuation or not? It'll be helpful if more information provided for reproducing this issue. Thanks. > /* >> * If the new TLB flushing is pending during flushing, leave >> * mm->tlb_flush_batched as is, to avoid losing flushing. >> >> there have a performance improvement with whole cores, above +30% > > But I don't think it is a proper patch. There is no guarantee the cpu calling > flush_tlb_batched_pending is exactly the cpu sending the deferred > tlbi. so the solution is unsafe. But since this temporary code can bring the > 30%+ performance improvement back for high concurrency, we have huge > potential to finally make it. > > Unfortunately I don't have an arm64 server to debug on this. I only have > 8 cores which are unlikely to reproduce regression which happens in > high concurrency with 96 parallel tasks. > > So I'd ask if @yicong or someone else working on kunpeng or other > arm64 servers is able to actually debug and figure out a proper > patch for this, then add the patch as 5/5 into this series? > >> >> ./Run -c 96 -i 1 shell1 >> 96 CPUs in system; running 96 parallel copies of tests >> >> Shell Scripts (1 concurrent) 109229.0 lpm (60.0 s, 1 samples) >> System Benchmarks Partial Index BASELINE RESULT INDEX >> Shell Scripts (1 concurrent) 42.4 109229.0 25761.6 >> ======== >> System Benchmarks Index Score (Partial Only) 25761.6 >> >> >> Tested-by: Xin Hao<xhao@xxxxxxxxxxxxxxxxx> > > Thanks for your testing! > >> >> Looking forward to your next version patch. >> >> On 7/11/22 11:46 AM, Barry Song wrote: >>> Though ARM64 has the hardware to do tlb shootdown, the hardware >>> broadcasting is not free. >>> A simplest micro benchmark shows even on snapdragon 888 with only >>> 8 cores, the overhead for ptep_clear_flush is huge even for paging >>> out one page mapped by only one process: >>> 5.36% a.out [kernel.kallsyms] [k] ptep_clear_flush >>> >>> While pages are mapped by multiple processes or HW has more CPUs, >>> the cost should become even higher due to the bad scalability of >>> tlb shootdown. >>> >>> The same benchmark can result in 16.99% CPU consumption on ARM64 >>> server with around 100 cores according to Yicong's test on patch >>> 4/4. >>> >>> This patchset leverages the existing BATCHED_UNMAP_TLB_FLUSH by >>> 1. only send tlbi instructions in the first stage - >>> arch_tlbbatch_add_mm() >>> 2. wait for the completion of tlbi by dsb while doing tlbbatch >>> sync in arch_tlbbatch_flush() >>> My testing on snapdragon shows the overhead of ptep_clear_flush >>> is removed by the patchset. The micro benchmark becomes 5% faster >>> even for one page mapped by single process on snapdragon 888. >>> >>> >>> -v2: >>> 1. Collected Yicong's test result on kunpeng920 ARM64 server; >>> 2. Removed the redundant vma parameter in arch_tlbbatch_add_mm() >>> according to the comments of Peter Zijlstra and Dave Hansen >>> 3. Added ARCH_HAS_MM_CPUMASK rather than checking if mm_cpumask >>> is empty according to the comments of Nadav Amit >>> >>> Thanks, Yicong, Peter, Dave and Nadav for your testing or reviewing >>> , and comments. >>> >>> -v1: >>> https://lore.kernel.org/lkml/20220707125242.425242-1-21cnbao@xxxxxxxxx/ >>> >>> Barry Song (4): >>> Revert "Documentation/features: mark BATCHED_UNMAP_TLB_FLUSH doesn't >>> apply to ARM64" >>> mm: rmap: Allow platforms without mm_cpumask to defer TLB flush >>> mm: rmap: Extend tlbbatch APIs to fit new platforms >>> arm64: support batched/deferred tlb shootdown during page reclamation >>> >>> Documentation/features/arch-support.txt | 1 - >>> .../features/vm/TLB/arch-support.txt | 2 +- >>> arch/arm/Kconfig | 1 + >>> arch/arm64/Kconfig | 1 + >>> arch/arm64/include/asm/tlbbatch.h | 12 ++++++++++ >>> arch/arm64/include/asm/tlbflush.h | 23 +++++++++++++++++-- >>> arch/loongarch/Kconfig | 1 + >>> arch/mips/Kconfig | 1 + >>> arch/openrisc/Kconfig | 1 + >>> arch/powerpc/Kconfig | 1 + >>> arch/riscv/Kconfig | 1 + >>> arch/s390/Kconfig | 1 + >>> arch/um/Kconfig | 1 + >>> arch/x86/Kconfig | 1 + >>> arch/x86/include/asm/tlbflush.h | 3 ++- >>> mm/Kconfig | 3 +++ >>> mm/rmap.c | 14 +++++++---- >>> 17 files changed, 59 insertions(+), 9 deletions(-) >>> create mode 100644 arch/arm64/include/asm/tlbbatch.h >>> >> -- >> Best Regards! >> Xin Hao >> > > Thanks > Barry > . >